Deep learning-based automated cervical length measurement improves spontaneous preterm birth risk stratification in women with a borderline short cervix (20-25 mm) at mid-trimester: A retrospective cohort study.

Authors

Abstract

This study evaluates whether deep learning (DL)-based automated cervical length (CL) measurement provides improved risk stratification for spontaneous preterm birth (PTB), compared with conventional CL measurement methods, in women with a borderline short cervix by conventional methods (CL 20-25 mm) on mid-trimester ultrasound. This retrospective cohort study was conducted in a tertiary referral hospital in South Korea between January 2019 and December 2023. Women with a conventionally measured borderline short cervix were reanalyzed using a previously validated DL-based algorithm (CL-Net) that traces the cervical canal, and were reclassified as DL-Normal (DL-CL > 25 mm) and DL-Short (DL-CL ≤ 25 mm) groups. The primary outcome was the risk of PTB before 37 weeks of gestation. Discrimination of conventional and DL-based CL was compared using receiver operating characteristic (ROC) curves with the DeLong test, and the incremental value of DL-based measurement was assessed by reclassification analyses. Of 171 women, 39 (22.8%) were reclassified into a DL-Short and 132 (77.2%) into a DL-Normal group. PTB before 37 weeks occurred in 28.2% of the DL-Short group and 6.1% of the DL-Normal group (P < 0.01), and a DL-Short cervix was independently associated with PTB (adjusted odds ratio, 6.37; 95% confidence interval, 2.28-17.84). DL-based CL showed better discrimination than conventional CL (area under the curve, 0.70 vs. 0.59; DeLong P = 0.042), with no evidence of poor calibration (Hosmer-Lemeshow, P = 0.36). The DL-Short group had a higher PTB rate despite greater progesterone use. For a DL-based CL ≤ 25 mm, sensitivity, specificity, positive predictive value, and negative predictive value for PTB were 57.9%, 81.6%, 28.2%, and 93.9%, respectively. Deep learning-based CL assessment, which follows the true anatomical pathway of the cervical canal, provides an objective and reproducible way to refine PTB risk stratification in women with a borderline short cervix. By identifying women at high risk of PTB more precisely than conventional methods, this approach could inform clinical practice, potentially helping to avoid unnecessary interventions in lower-risk women while ensuring appropriate preventive care for at risk patients with a borderline short cervix.

Tags

Topics